From: Mike Galbraith <mgalbraith@suse.de>
Date: Wed Aug  2 10:55:41 CEST 2017
Subject: sched: Mitigate LB impact on RT
Patch-mainline: Never, RT specific
References: SLE Realtime Extension

Signed-off-by: Mike Galbraith <mgalbraith@suse.de>
---
 kernel/sched/fair.c     |   15 +++++++++++++--
 kernel/sched/features.h |    4 ++++
 kernel/sched/sched.h    |   10 ++++++++++
 3 files changed, 27 insertions(+), 2 deletions(-)

--- a/kernel/sched/fair.c
+++ b/kernel/sched/fair.c
@@ -6327,6 +6327,10 @@ pick_next_task_fair(struct rq *rq, struc
 	return p;
 
 idle:
+	/* RT tasks have better things to do than play load balancer */
+	if (IS_ENABLED(CONFIG_PREEMPT_RT_FULL) && prev->sched_class != &fair_sched_class)
+		return NULL;
+
 	new_tasks = idle_balance(rq, rf);
 
 	/*
@@ -6841,9 +6845,11 @@ static int detach_tasks(struct lb_env *e
 		/*
 		 * NEWIDLE balancing is a source of latency, so preemptible
 		 * kernels will stop after the first task is detached to minimize
-		 * the critical section.
+		 * the critical section.  For PREEMPT_RT_FULL, minimize all
+		 * load balancing perturbation, it's better to nibble frequently
+		 * than to take big gulps.
 		 */
-		if (env->idle == CPU_NEWLY_IDLE)
+		if (IS_ENABLED(CONFIG_PREEMPT_RT_FULL) || env->idle == CPU_NEWLY_IDLE)
 			break;
 #endif
 
@@ -8082,7 +8088,12 @@ static int load_balance(int this_cpu, st
 		env.loop_max  = min(sysctl_sched_nr_migrate, busiest->nr_running);
 
 more_balance:
+#ifdef CONFIG_PREEMPT_RT_FULL
+		if (!rq_lock_trylock_irqsave(busiest, &rf))
+			goto out_balanced;
+#else
 		rq_lock_irqsave(busiest, &rf);
+#endif
 		update_rq_clock(busiest);
 
 		/*
--- a/kernel/sched/features.h
+++ b/kernel/sched/features.h
@@ -86,7 +86,11 @@ SCHED_FEAT(RT_PUSH_IPI, true)
 #endif
 
 SCHED_FEAT(RT_RUNTIME_SHARE, true)
+#ifdef CONFIG_PREEMPT_RT_FULL
+SCHED_FEAT(LB_MIN, true)
+#else
 SCHED_FEAT(LB_MIN, false)
+#endif
 SCHED_FEAT(ATTACH_AGE_LOAD, true)
 
 SCHED_FEAT(WA_IDLE, true)
--- a/kernel/sched/sched.h
+++ b/kernel/sched/sched.h
@@ -1670,6 +1670,16 @@ rq_lock_irqsave(struct rq *rq, struct rq
 	rq_pin_lock(rq, rf);
 }
 
+static inline int
+rq_lock_trylock_irqsave(struct rq *rq, struct rq_flags *rf)
+	__acquires(rq->lock)
+{
+	if (!raw_spin_trylock_irqsave(&rq->lock, rf->flags))
+		return 0;
+	rq_pin_lock(rq, rf);
+	return 1;
+}
+
 static inline void
 rq_lock_irq(struct rq *rq, struct rq_flags *rf)
 	__acquires(rq->lock)